WPS-AI

# Lingxi AI · WPS Office Multi-Host AI Assistant **🌐 Language**: English · [English](README.en.md) A TaskPane compatible with **WPS Writer / Spreadsheets / Presentation / PDF** four endpoints, supporting multiple AI providers (Codex / OpenAI compatible / Anthropic) + one image provider. AI can directly read and write documents through tools. > 🤖 100% completed with Claude vibe coding: architecture, provider adaptation, PPT themes/charts, Word rendering, cross-platform installer, and documentation through [Claude](https://claude.com/claude-code) and human conversation iteration. No manual coding in the repository - feel free to fork and vibe. --- ## Screenshots | WPS Writer | WPS Spreadsheets | WPS Presentation | |---|---|---| |  |  |  | --- ## Get Started in 5 Minutes ### 1. Download Installation Package | Platform | Installation Package | Size | |---|---|---| | **Windows** | [`lingxi-ai-1.3.0-setup.exe`](https://github.com/lewis-hui1202/WPS-AI/releases) | ~30 MB | | **macOS** | [`lingxi-ai-1.3.0.pkg`](https://github.com/lewis-hui1202/WPS-AI/releases) (double-click to install) | ~35 MB | Both packages include a built-in Node runtime, **no need to install Node separately**. ### 2. Installation - **Windows**: Temporarily disable antivirus real-time protection → double-click setup.exe → completely exit WPS → restart WPS, and the "Lingxi AI" tab will appear in the ribbon. - **macOS**: **Right-click .pkg → Open** (Gatekeeper may block unsigned packages, double-click may result in an error) → enter system password → completely exit WPS → restart WPS. Detailed steps (uninstall / upgrade / troubleshooting / installer building) see [INSTALL.md](INSTALL.md). ### 3. Configure AI Model 1. Click "Open Lingxi AI" on the ribbon → TaskPane pops up on the right → ⚙ settings (independent 960×720 popup). 2. "Chat Model" panel → **+ Add Vendor** → select one of the 12 presets (baseURL is pre-filled). 3. Enter API Key → ⚡ test → close popup → select model from header dropdown to start chatting. **Multiple providers can be used simultaneously**: DeepSeek + Anthropic + Codex + Kimi can be used together, and switched from the dropdown. | Preset | baseURL (pre-filled) | |---|---| | **Codex (ChatGPT OAuth)** | (OAuth) | | **Anthropic** / **OpenAI** | `api.anthropic.com` / `api.openai.com` | | **DeepSeek** / **Kimi** / **Qwen** | DeepSeek / Moonshot / DashScope compatible addresses | | **GLM** / **豆包** / **硅基** / **OpenRouter** | OpenAI compatible protocol addresses | | **Local Ollama** | `http://localhost:11434/v1` (no API key required) | | **Custom** | URL + Key self-filled | #### System Requirements | Item | Windows | macOS | |---|---|---| | **Operating System** | Windows 10 / 11 (x64) | macOS 10.15 Catalina+ (Intel + Apple Silicon) | | **WPS** | WPS Office 12.x+ | WPS Office 5.x+ | | **Runtime** | Built-in portable Node 22.x | Built-in darwin-x64 + arm64 Node | | **Permissions** | Default installation to user directory | Installation requires system password (writing to `/Library/Application Support/LingxiAI/`) | Not supported: WebOffice, mobile WPS, low-version desktop clients — JSAPI plugin requires desktop clients + above minimum version. --- ## Feature Overview ### Common Features **AI Integration** - Support multiple chat providers (Codex / Anthropic / OpenAI compatible 11 providers + custom), header dropdown grouped by provider. - Three ability icons next to the model: 🖼 Image / 📄 PDF / 💡 Thinking. - Thinking intensity adjustable (low / medium / high / off), mapped to `thinking.budget_tokens` / `reasoning_effort` / `reasoning.effort` by provider. - PDF as multimodal attachment feeding large models, Claude uses document block / OpenAI uses Files API / Codex uses input_file. - Streaming output + tool-use loop: continuously call multiple tool operations within a single conversation. - **Preview confirmation mode** vs **AI direct writing mode**: two safety levels. **TaskPane Experience** - Settings / PPT style presets independent WPS dialog window, detached from TaskPane width limit. - One-click detach / dock: floating window with 8 corner resize handles. - Popups adapt to screen resolution (no overflow on small screens). **Conversation** - Multi-conversation management + complete history replay (inference + tool calls + text response). - Tool call bubble similar to Claude Code: default shows only tail parameter preview, result arrives and disappears; can open developer log in settings to view complete JSON. - Progress bar with recent output segment. - AI working during document lock banner + progress merge (no visual overlap). - Attachment support: image / PDF / text file. - System prompt customizable; support "skills" (built-in 4 sets + import .md/.txt) to拼 to system prompt. **Operation Security** - AI working during hard document lock (Word `Document.Protect` / Excel `UserInterfaceOnly`). - Temporary document reject modification (before chatting, fail-fast to avoid waiting for AI output). - Per-turn document backup + one-click rollback (automatic GC recent 20 documents). - Modification record Tab: grouped by file, expand to view input/before snapshot/error. - Configuration import and export (API Key encryption + version compatibility). ### Writer / Spreadsheets / Presentation / PDF | Host | Feature Overview | |---|---| | **Writer** | 6 groups of shortcut buttons (writing / rewriting / polishing / translation / summarization / intelligent) + markdown→Word native format (true table + nested list + paragraph indent forced to zero) | | **Spreadsheets** | Cell/range read/write, batch formatting, table beautification, column width adaptive, AI-generated formula / conversion table / proofreading | | **Presentation** | 50+ sets of color palettes with design concept + 8 shape templates + 4 SVG visual templates + 6 data visualization charts + PPT outline generation + HTML template system (17 layouts: cover / section / content / stat / feature-grid / quote / comparison / metric-trio / timeline / agenda / two-column / image-text / process / table / bento / closer / freeform) + visual editor (drag / 8-direction resize / multi-select / PS-style alignment guides + snap) + ECharts + unified style + AI flavor removal | | **PDF** | Side-by-side translation (markdown original/translation table) + full-text summarization + document-generated PPT outline + PDF Q&A + intelligent recommendation operation | ### MCP Service (added in v1.4) Settings → MCP Service → enable exposes WPS tools to external agents (Claude Code CLI / Claude Desktop / Cursor, etc.): - JSON configuration one-click copy (automatically fill in plugin installation path). - Tools displayed grouped by host. - Real-time status badge (connected / enabled but not connected / disabled). --- ## Project Structure ```text plugin/ ├── taskpane.html # Business UI entry ├── main.js # Script loader ├── manifest.json / ribbon.xml # Plugin declaration + ribbon ├── css/style.css ├── js/ │ ├── app.js # Business UI orchestration │ ├── wps.js / hosts/* # Host distribution + host jsapi bridge │ ├── providers/* # Provider registry + 12 presets + capabilities │ ├── tools/* # AI-callable tools (grouped by host + registry) │ ├── html-templates/* # Template system (cache / components / renderer / studio) │ ├── mcp-bridge.js # MCP bridge (plugin side) │ └── skills.js # Skills (built-in + imported) └── tools/ ├── proxy-server.js # CORS proxy + file / backup / MCP endpoint ├── mcp-server.js # stdio MCP server (for external agents) ├── serve-permanent.js # Permanent mode static server ├── build-variants.js # Multi-host variant packaging └── dev.js / gen-ribbon.js ``` --- ## Secondary Development ```bash cd plugin npm install npm run dev:wps # or dev:et / dev:wpp / dev:pdf ``` Simultaneously starts CORS proxy (3890) and wpsjs debug (3889), WPS automatically awakens. **Add new tool** — edit `js/tools/<host>.js`: ```js registry.registerTool({ name: "wps_my_tool", hosts: ["wps"], // or ["wps","et","wpp","pdf"] description: "...", // The clearer AI knows when to call parameters: { type: "object", properties: { ... }, required: [...] }, handler: async (params) => ({ ok: true, ... }) }); ``` **Add new ribbon button** — edit `js/quick-actions.js`, run `npm run gen-ribbon`. **Permanent mode packaging** — `node tools/build-variants.js --out C:/path/dist-permanent`. --- ## Known Limitations - Exclusive to WPS desktop clients, Web/Mobile WPS not supported. - Mac WPS WKWebView permanent mode occasionally requires clearing WebKit cache after reinstallation (see [INSTALL.md](INSTALL.md) Q7). - wpsjs debug can only register one host at a time, debugging switch host mainly restarts. - Codex OAuth is a non-official reuse solution; OpenAI adjusting authorization strategy may require updating client_id. - Image provider protocol bound toapis; switching to other services requires adapting in `providers/image.js`. --- ## Feedback ### Join Fan Group Scan the QR code to follow the official account, reply `ai` to get the fan group link, and communicate with other users or report bugs: <img src="img/qrcode_for_gh_e26e731fb54c_258.jpg" alt="Official Account QR Code" width="160" /> > WeChat scan → follow official account → reply `ai` → join fan group ### Bug Report Information If you find a bug, please report in the group or [GitHub Issues](https://github.com/lewis-hui1202/WPS-AI/issues) with: - System + WPS version - Which host, which tool triggered - Console error (TaskPane internal right-click → inspect / "Open JS Debugger") - Last 50 lines of `~/.lingxi-ai/server.log` (permanent mode) --- ## Update Log Complete change record see [CHANGELOG.md](CHANGELOG.md). - **v1.4** (completed): HTML template system expanded to 17 layouts (cover/section/content/stat/feature-grid/quote/comparison/metric-trio/timeline/agenda/two-column/image-text/process/table/bento/closer/freeform, typography + editing feel details + golden ratio transformation) / HTML preview visual editor (arbitrary layout automatically converted to freeform, then drag / 8-direction resize / multi-select / PS-style alignment guides + 6px snap + Shift temporary disable / coordinate hint) / skills (opt-in, no default enable) / MCP service / responsive popup / PPT style preset independent window + visual theme grid + real-time preview / grayscale hot update (device SN + canary whitelist + rolloutPercent percentage release) / local Ollama model selection suggestion / image generation (toapis + codex-bridge, including CF JA3 interception accurate error attribution) - **v1.3**: PDF host / multi-vendor management / settings popup / TaskPane free layout - **v1.2.1**: multi-conversation management / modification record / document lock / configuration encryption - **v1.1**: permanent installation upgrade / Mac WKWebView cache fix / IME input compatibility - **v1.0**: first version (permanent installation / 12 themes / 6 charts / multi-provider)